I just stumbled over a good comment from Jonathan Ellis at http://www.datastax.com/dev/blog/cql3-for-cassandra-experts
The rule of thumb is, use one columnfamily for each type of resultset you want to query. Cassandra is not Bigtable or HBase where you have to try to cram everything into very few ColumnFamilies because of implementation limitations.
As a bonus, “one CF per resultset type” promotes good performance, because any time you’re selecting from the beginning of the row, you will get better performance than if you have to scan past other cells.
Calloc.org 